Instructor: Vincent Conitzer CPS 270: Artificial Intelligence http://www.cs.duke.edu/courses/fall08/cps270/ Markov decision processes, POMDPs Instructor: Vincent Conitzer

Warmup: a Markov process with rewards We derive some reward from the weather each day, but cannot influence it .6 s 10 .1 .3 .4 .3 8 .2 1 c r .3 .3 .5 How much utility can we expect in the long run? Depends on discount factor δ Depends on initial state

Figuring out long-term rewards Let v(s) be the (long-term) expected utility from being in state s now Let P(s, s’) be the transition probability from s to s’ We must have: for all s, v(s) = R(s) + δΣs’ P(s, s’) v(s’) .6 s 10 .1 .3 .4 .3 8 .2 1 c r .3 .3 .5 E.g., v(c) = 8 + δ(.4v(s) + .3v(c) + .3v(r)) Solve system of linear equations to obtain values for all states

Iteratively updating values If we do not want to solve system of equations… E.g., too many states … can iteratively update values until convergence vi(s) is value estimate after i iterations vi(s) = R(s) + δΣs’ P(s, s’) vi-1(s’) Will converge to right values If we initialize v0=0 everywhere, then vi(s) is expected utility with only i steps left (finite horizon) Dynamic program from the future to the present Shows why we get convergence: due to discounting far future does not contribute much

Markov decision process (MDP) Like a Markov process, except every round we make a decision Transition probabilities depend on actions taken P(St+1 = s’ | St = s, At = a) = P(s, a, s’) Rewards for every state, action pair R(St = s, At = a) = R(s, a) Sometimes people just use R(s); R(s, a) little more convenient sometimes Discount factor δ

Example MDP Machine can be in one of three states: good, deteriorating, broken Can take two actions: maintain, ignore

Policies No time period is different from the others Optimal thing to do in state s should not depend on time period … because of infinite horizon With finite horizon, don’t want to maintain machine in last period A policy is a function π from states to actions Example policy: π(good shape) = ignore, π(deteriorating) = ignore, π(broken) = maintain

Evaluating a policy Key observation: MDP + policy = Markov process with rewards Already know how to evaluate Markov process with rewards: system of linear equations Gives algorithm for finding optimal policy: try every possible policy, evaluate Terribly inefficient

Bellman equation Suppose you are in state s, and you play optimally from there on This leads to expected value v*(s) Bellman equation: v*(s) = maxa R(s, a) + δΣs’ P(s, a, s’) v*(s’) Given v*, finding optimal policy is easy

Value iteration algorithm for finding optimal policy Iteratively update values for states using Bellman equation vi(s) is our estimate of value of state s after i updates vi+1(s) = maxa R(s, a) + δΣs’ P(s, a, s’) vi(s’) Will converge If we initialize v0=0 everywhere, then vi(s) is optimal expected utility with only i steps left (finite horizon) Again, dynamic program from the future to the present

Policy iteration algorithm for finding optimal policy Easy to compute values given a policy No max operator Alternate between evaluating policy and updating policy: Solve for function vi based onπi πi+1(s) = arg maxa R(s, a) +δΣs’ P(s, a, s’) vi(s’) Will converge

Mixing things up Do not need to update every state every time Makes sense to focus on states where we will spend most of our time In policy iteration, may not make sense to compute state values exactly Will soon change policy anyway Just use some value iteration updates (with fixed policy, as we did earlier) Being flexible leads to faster solutions

Linear programming approach If only v*(s) = maxa R(s, a) + δΣs’ P(s, s’, a) v*(s’) were linear in the v*(s)… But we can do it as follows: Minimize Σs v(s) Subject to, for all s and a, v(s) ≥ R(s, a) + δΣs’ P(s, s’, a) v(s’) Solver will try to push down the v(s) as far as possible, so that constraints are tight for optimal actions

Partially observable Markov decision processes (POMDPs) Markov process + partial observability = HMM Markov process + actions = MDP Markov process + partial observability + actions = HMM + actions = MDP + partial observability = POMDP full observability partial observability Markov process HMM MDP POMDP no actions actions

Example POMDP Need to specify observations E.g., does machine fail on a single job? P(fail | good shape) = .1, P(fail | deteriorating) = .2, P(fail | broken) = .9 Can also let probabilities depend on action taken

Optimal policies in POMDPs Cannot simply useπ(s) because we do not know s We can maintain a probability distribution over s using filtering: P(St | A1= a1, O1= o1, …, At-1= at-1, Ot-1= ot-1) This gives a belief state b where b(s) is our current probability for s Key observation: policy only needs to depend on b, π(b)

Solving a POMDP as an MDP on belief states If we think of the belief state as the state, then the state is observable and we have an MDP (.3, .4, .3) observe failure disclaimer: did not actually calculate these numbers… maintain observe success (.5, .3, .2) (.6, .3, .1) (.2, .2, .6) ignore Reward for an action from a state = expected reward given belief state observe failure observe success (.4, .2, .2) Now have a large, continuous belief state… Much more difficult